Sam Altman just confirmed that OpenAI will "definitely" build humanoid robots, envisioning personal robots for everyone down the line. Job postings show the company is already assembling teams to manage data collection facilities and robot hardware. Why it matters: Moving beyond software puts OpenAI into direct hardware competition with Figure and Tesla while securing proprietary real-world training data.
Hugging Face’s Pollen Robotics just unveiled Microduck, a sub-10-inch bipedal robot available for preorder at $399. The open-source bot can rollerskate, retrieve items, and follow laser pointers, all while generating a permanent unique voice on its very first boot. It is basically an affordable, hackable robotic pet built for both play and AI experimentation.
Skild AI just launched S1, a robotics foundation model that learns unseen, 10-minute tasks from a single video prompt without any fine-tuning. Using in-context learning similar to LLMs, it matches conventional VLA models that typically require 50 to 100 hours of training data. It is currently being rolled out to select industrial partners.
Generalist AI just unveiled GEN-1.5, a robot foundation model that learns new physical tasks from a single 3-to-12-second demonstration without any gradient updates or fine-tuning. Across 10 manipulation tasks, it hit an average 59% success rate via in-context physical prompting, rising to 83% with five minutes of data. It basically brings GPT-3-style in-context learning to real-world robotics, though behaviors remain short-horizon for now.
Unitree just teased a new humanoid robot nicknamed "Superman" that can hit a top speed of 12.66 m/s and jump 2 meters high from a standing start. With 0.85-meter legs and just over three months in development, the company claims it already beats all human records, with more performance gains expected soon.
PaXini AI just launched the PXCap Pro, a 300g data collection glove designed to capture high-quality human hand demonstrations for robot training. It packs tactile sensors sensitive down to 0.1 Newtons, a wrist-mounted 165° wide-angle camera, and magnetic-resistant joint encoders. By outputting standardized physical data straight from hardware, it tackles the real bottleneck of teaching robots delicate tasks like packaging or lab handling.
Chinese startup Arkshel Robotics just revealed a shapeshifting robot prototype called MX01. The machine can switch between a bipedal humanoid and a four-legged walker, then dock with an industrial drone module for aerial transport, with a wheeled mode in the works. Why it matters: it offers extreme terrain flexibility for complex missions, though packing so many configurations into one machine inevitably multiplies mechanical points of failure.
Dyna Robotics just unveiled DYNA-2, a world-action model trained on over 1 million hours of first-person human video with zero robot data in pre-training. By watching how objects react to human hands, the system builds physical intuition and transfers to real robot hardware after just a few hours of fine-tuning. It shows robots can skip tedious teleoperation data and learn directly from the way we move.
Boston Dynamics' Atlas humanoid robot can now navigate back to a charging station and swap its own depleted battery in less than three minutes. Standard charging takes around 90 minutes for a four-hour runtime, so quick autonomous swaps remove a major bottleneck for running non-stop factory shifts.
Google DeepMind just launched Gemini Robotics ER 2, an embodied reasoning model designed to act as a high-level brain for robots. It tracks tasks via live video with 91.3% moment-finding accuracy, plans next steps while moving, and coordinates multiple robots at once. Why it matters: it eliminates the awkward "stop-and-think" pauses that keep physical AI stuck in labs.
The Trump administration just directed the FCC to ban imports of new Chinese humanoid and quadruped robots, alongside connected power inverters. Officials claim the hardware poses cybersecurity and data theft risks to the U.S. AI buildout. The move aims to wall off critical data center supply chains while forcing manufacturing back stateside.
Lunar Outpost just announced its next moon rover will use Nvidia Jetson chips to run its lidar system, potentially putting the first GPU on the lunar surface. Slated to launch on a Falcon 9 rocket before the end of the year, the mission will test if commercial physical AI chips can survive cosmic radiation and freezing lunar nights on minimal power.
Black Forest Labs just teamed up with mimic robotics to turn its new FLUX 3 multimodal foundation model into FLUX-mimic, a video-action AI running on robots in Audi factories. By tapping into the physics FLUX learns during video generation, a lightweight decoder translates scene dynamics into physical manipulation in under 80ms on a single RTX 5090.
Spatial intelligence startup World Labs just acquired robotics company SceniX to bring its models into physical systems. The team argues the next robotics breakthrough relies on combining spatial intelligence, world models, and real-world learning loops. Detailed capabilities from SceniX will be shared soon.
Unitree just demonstrated UnifoLM-OminiA-0.3, an omni-modal AI model designed to handle real-time home-care tasks through whole-body mobile manipulation. The autonomous system can interpret multimodal inputs while resisting physical disturbances, successfully performing chores like adjusting a hospital bed. It is a tangible step toward putting robotic caregivers into real homes.
Nvidia and Stanford researchers just unveiled RoboTTT, scaling a robot's visuomotor context to 8,000 timesteps—roughly five minutes of continuous real-time control. By using Test-Time Training to compress action history into fast weights without adding inference latency, the robot aced a 10-stage assembly task where traditional baselines failed. Why it matters: Context length is emerging as the next big scaling law for physical AI.
